The lines that matter, and what each one means
Lines from a real report: the morning move on NQ, the shelf’s own test, run November 1, 2026, on data to October 30, 2026 (its page, where the whole report is folded in). Press "why?" on any of them. A failed report has the same lines; only the numbers and the verdict change.
SCORE 90 / 100 VERDICT PASSED
Six checks, added up. Seventy or more, with no check under half its points, no automatic fail, and two more checks (the coin flips at 1 in 20 counting every try; the range of outcomes on the years before 2023 above zero), is a pass.
Think of it like a safety inspection with six stations. A car can’t pass by acing the brakes if the steering fails. Passed means the strategy cleared every station reasonably and most of them well, and then cleared two more: it beat the coin flips, and its range of likely results on the years before 2023 never crosses zero. Close is 40 to 69, or a 70 that missed one of those two; when it is the second kind, the report says which. Failed means the written steps did not beat a coin, or something stopped the test whatever the score: too few trades, money that only came since 2023, or money made before 2023 and lost since. Six checks: clean data 5, the years before 2023 holding up 25, year by year 20, the coin flips 15, costs 20, settings nudged 15. Passed: 70 or more, no check under half its points, no stop, and since October 25, 2026 the two more checks (the coin flips at 1 in 20 counting every try; the bootstrap range on the years before 2023 above zero); a 70 that misses one is Close. Close otherwise is 40 to 69; Failed is under 40 or a stop (too few trades, look-ahead, a reproduction failure, money only since 2023). The three checks worth 60 of the 100 points are scored on the years before 2023; the other three use all 18.Before 2023 p = 0.0030 (15 / 5000) <- governs
On the years before 2023, only 15 of 5,000 coin flips did as well as this strategy.
Five thousand copies of the strategy take its own trades on the same days with the side picked at random (for a one-way setup, the timing). The smaller this number, the less likely the result was luck. 0.05 means one coin flip in twenty matched it, which is the bar a pass needs. 0.003 means three in a thousand. Above 0.20 or so, the strategy is a coin on those days. The flips run on 2008 to 2022, the years kept apart from the ones since; the full-window line above it is printed and not used. Empirical p: the share of 5,000 random-side copies with net ≥ the real net, on the trades before 2023 only (the seen-from year is declared on the form; 2023 on the shelf). The count is reported, not a fitted z; the tail is fatter than a normal approximation would give. The line that counts every try follows.Counting earlier tries, p (try #1) = 0.0030
This was the first version tried, so the odds stand as they are. Say it was the third, and this line becomes 0.0090.
If you roll a die three times and keep the best roll, the best roll isn’t a fair picture of the die. Same with strategies. The more versions you tried, the more the winner is discounted. This line does that arithmetic with the number you gave on the form, and every condition you add counts as one more. It is this line, not the raw one, that the pass and the code file read. Multiplicity: 1 − (1 − p)n, n the declared try (the larger of what you declared and the submission count on record for your address). Šidák-style; conservative for correlated versions and the best available correction when the search history is self-reported. 1 − 0.9973 = 0.0090.Pre-2023 CI $12,788.20 .. $118,338.20 P(net>0) 0.9924
Resample the trades before 2023 five thousand times, drawing with replacement, and 95 in 100 of them land between $12,788 and $118,338; it is above zero in 99.2% of the resamples.
The second of the two extra checks a pass needs. Draw the same number of trades again from its own list, some twice and some not at all, five thousand times: this is the range the net lands in. If the low end sits above zero, the average trade was above zero by more than the trades’ own spread, taking each trade as independent; a strategy whose money sits in one or two years can pass this and still depend on those years. If the range crosses zero, a 70 or better is held at Close, and the report says so. Bootstrap × 5,000 on the trade list before 2023 (resampled with replacement), 95% interval of net; the full-window interval is printed above it. The pass needs the pre-2023 interval to exclude zero. DD95/actual from the same shuffles feeds the coin-flips check (big trades).Largest year 2022 = 44.7% of the net before 2023
Almost half of what it made on the years before 2023 came from one year.
A strategy can be real and still be one you’d never have stuck with. If most of the money comes from one wild year, you’d have spent the other years wondering why you bother, and most people quit before the good year shows up. This line is the difference between “it works” and “you could have traded it.” Year by year: each calendar year before 2023 a window; the share profitable (12 of 15 here), the profit factor before 2023 against since, and the largest single year’s share of the net before 2023 (over 40% costs points). Beside it: top trade and top-5 share, the share of rolling 12-month windows negative, and DD95/actual.Breakeven 16.9 ticks/side
Slippage would have to reach 17 ticks on every entry and exit before the strategy stops making money.
Every time you trade, you lose a little to the gap between buying and selling prices. One tick is normal; three is a rough day. A strategy that breaks even at 17 ticks has a lot of room. A strategy that breaks even at 2 ticks works on paper and not in your account. It feels free; the report shows what it cost. Searched by full re-simulation at integer ticks a side with linear interpolation between rungs; for bracketed strategies the exits are re-resolved at each rung. The same section re-runs the strategy at 2 and 3 ticks everywhere, and prints where costs bite: the trades that start or end within 5 minutes of an 8:30 or 2:00 release, or start outside 9:30 to 4:00, at 3 ticks a side. Under 4 ticks fails the code offer.Longest stretch without a new equity high: 46 months (ending 2014-09)
From 2010 into 2014 this strategy made no new high. Real, and slow.
This is the line that decides whether you could have held it. A strategy can pass every test and still spend four years going nowhere, and most people quit in year two. Read the number as the wait you would have had to sit through. It is printed and not scored, on purpose: a real-but-slow edge and a fake-but-smooth one must never look alike. Section 9, no points: the longest stretch in months between one equity high and the next on the full trade list at one contract; the share of rolling 12-month windows negative; the worst 12 months; the maximum drawdown and the years of the median year’s net it takes to recoup it (6.2 here); the share of all profit from the best two years.Long 363 $98,935 · Short 353 $14,532 · Wednesday 147 −$4,332
Where it breaks: the longs made seven times what the shorts did, and Wednesdays lost. Every side and every weekday is in the table.
The same trades, split by side and by the day they were taken, each with its count, its net and its winners. A split is something to look at, not a new rule: keeping only the best side or day is a new version, and testing it counts as one more try. On a result page each split is one tap away from the form with that one change filled in; in the mail, the links are under this table. Section 3, “How it splits”: Long, Short and each weekday, with trades, net, average and winners, over every trade at one contract after costs. Also printed, from the autopsy: what was on the books at each hour before 2023, and how each exit type did.No single change is pointed at: the verdict stands as written. Where it breaks, by side and by weekday, is under How it splits (3. RESULT); each is one change to test as a second version, scored with a tighter bar.
The report says what it would run next, or that nothing is pointed at and why. Here, nothing: the strategy passed as written.
On a failed report this line usually names one change and the prediction for it, written before anyone runs it: “The one next run, if you want it: the clock exit at 12:00, as try #2. Prediction on file: …” Three such runs at most, then the verdict stands. On the shelf, that run has been made for every setup and its result sits on the page under “What we ran next”. next_run: five rules in order (a stop that fires too often, a condition that subtracts, the clock exit that already held three quarters of the money, an era that no change fixes, a plateau under 25%); otherwise nothing is pointed at. The prediction is filed in the record before the run; the next run is scored as the declared try + 1 and its result is compared to the prediction on the result page.CODE OFFER => OFFERED
Five checks cleared, so this strategy came with a Python file you can run yourself. Not every kind of strategy gets one yet; when it doesn’t, the report says why.
If a strategy passes, beats the coin flips on the years before 2023 even after every try is counted, keeps its range of outcomes above zero there, has room for slippage, and doesn’t depend on one trade, we hand you the strategy as a program. It’s the same code we used to test it, so what you run is what we scored. It is not a trading robot and it doesn’t place orders. The five lines as the report prints them: passed; fewer than 1 in 20 coin flips matched it on the years before 2023, counting earlier tries; the range of outcomes on the years before 2023 stays above zero; it breaks even at 4 ticks of cost a side or more; no single trade is over 25% of the net. The file lifts the exact signal and simulation functions from the engine and is verified against it on the same data to the cent before delivery. Standard library only.